An extension of paper Learning Barrier Certificates for Neural Path Tracking Control of Self Driving Vehicles is contained in the PDF file, with the following sections
- Pseudocodes
- Policy Learning
- Learning Low-Dimensional Barriers under Partial Observability
- Estimating Range of Dynamics near Samples
- Finding Boundary Counterexamples for Retraining
- Using the Learned Barrier Function for Safety Monitor
- Hyper-parameters and Details of Experiments
Below we present figures contained in the original and extended paper.
Plotting of trajectories of the dynamic model, with x, y axes as angle and distance errors, and z axis as:
Longitudinal Speed | Lateral Speed | Yaw Rate |
---|---|---|
Kinematic Model | Dynamic Model |
---|---|
The barrier functions on kinematic model, dynamic model and TORCS environment
Kinematic Model (3D) | Dynamic Model (2D) | TORCS (2D) |
---|---|---|
Backward reachable states for dynamic model, evaluated on a maximum curvature of 0.15, for 50 time steps
Path with Curvature 0.15 (Curve to Left) | Path with Curvature 0.15 (Curve to Right) | Safety Monitor |
---|---|---|
The barrier function for dynamic model obtained after the initial training and after the final retraining, projected in the dimensions of angle and distance error
Initial Barrier | Final Barrier |
---|---|
Close to Collected Trajectories | Whole Certification Grid |
---|---|